Predicting the way something will smell is incredibly complex. That’s because we perceive scents using the millions of sensory neurons in our brains—and because scents are multi-faceted. For example, how would you describe the smell of an orange? Sweet? Bright? Grassy? Each descriptor is unique. Classifying smell is tricky because it requires a multi-label system.

Researchers at Google are building graph neural networks—a particular type of deep neural network that operates on graphs as inputs—to predict odors at a molecular level.